Artificial neural network based chemical mechanisms for computationally efficient modeling of hydrogen/carbon monoxide/kerosene combustion

Abstract

The high cost of high-resolution computational fluid/flame dynamics (CFD) has hindered its application in combustion related research, design and optimization. In this study, we propose a framework for turbulent combustion simulation based on the deep learning approach. An optimized deep convolutional neural network (CNN) inspired from a U-Net architecture and inception module is designed for constructing the deep learning solver, named CFDNN. CFDNN is then trained on simulation results of hydrogen combustion in a cavity with different inlet velocities. After training, CFDNN can not only accurately predict the flow and combustion fields within the range of the training set, but also shows an extrapolation ability for prediction outside the training set. The results from the CFDNN solver show excellent consistency with the conventional CFD results in terms of both predicted spatial distributions and temporal dynamics. Meanwhile, two orders of magnitude of acceleration are achieved by using the CFDNN solver compared to a conventional CFD solver. The successful development of such a deep learning-based solver opens up new possibilities of low-cost, high-accuracy simulations, fast prototyping, design optimization and real-time control of combustion systems such as gas turbines and scramjets.

Modeling of turbulent flames with the large eddy simulation–probability density function (LES–PDF) approach, stochastic fields, and artificial neural networks

“…The results in Figures 10-13 show that using individual networks for species rather than for composition subdomains has allowed a much smaller number of weights to be targeted at species predictions more effectively: this work uses approximately 30,000 weights for a mechanism of 31 species whereas Franke et al 36 used approximately 750,000 for a mechanism of 16 species, An et al 37 used approximately 4,500,000 for a mechanism of 41 species and Wan et al 39 used approximately 180,000 for a mechanism of 11 species.…”

“…39 ) or large numbers of medium sized ones (such as those in Refs. 36,37 ) meaning approaches based on stochastic gradient descent have to be used instead. This suggests that the number of MLP weights should be kept small wherever possible to allow more sophisticated training algorithms to be used.…”

“…Small, fast, MLPs can be used with high accuracy if the problem can be simplified. This simplification has previously been performed using subdivision of the composition space, with the SOM frequently used 28,[35][36][37] .…”

“…PDF simulation of the DLR flames A and B, and good agreement between DI and the SOM-MLP method was obtained.The SOM-MLP method was further developed by Franke et al36 to include flamelets with high strain rates to capture extinction, before application to an LES-PDF simulation of Sydney flame L, again with good agreement between the SOM-MLP method and DI. The SOM-MLP method was applied by An et al37 to the LES simulation of a rocket-based H 2 /CO/kerosene combustor, with training data collected via a RANS simulation of the same configuration rather than via unsteady flamelets. Other methods employed for generating ANN training data include the eddy dissipation model38 and a stochastic micromixing system39 .…”

Acceleration of Chemical Kinetics Computation with the Learned Intelligent Tabulation (LIT) Method

“…They used an abstract problem to generate the ground-truth data. Most recently, An et al [24] applied a similar method, called SOM-BPNN (Back-Propagation Neural Network), to model the combustion chemistry of hydrogen/hydrocarbon-fueled supersonic engines. Also, Owoyele and Pal [25] introduced a different data-driven tabulation method using a neural ODE approach, in which they have one network for each species of a H 2 -O 2 combustion.…”